What Predicted Attention Maps Can and Cannot Tell You
Predicted attention maps, also called saliency maps, estimate where eyes go from visual features like contrast, color, faces, and text. On mobile interfaces, data-driven models improved from an NSS score of 0.66 to 0.84, but only when they were trained on interface data (ACM MobileHCI, 2020). These maps measure visual pull. They do not measure what visitors read, want, or do.
What the research observed
Saliency models work bottom-up. They score what stands out visually. Real looking is also driven by the task and by expectation. On mobile interfaces, real gaze showed a strong top-left bias from expectation, and color and size mattered less than the models assumed. Classic parameter-free saliency models fit poorly (ACM MobileHCI, 2020).
How good a model looks also depends on how it is scored. Different evaluation metrics rank the same models differently because they treat errors, viewing biases, and spatial tolerance differently (IEEE TPAMI, 2019).
These models also cannot account for banner blindness, the learned habit of ignoring anything that looks like an ad (Nielsen Norman Group, 2018). A bright promotional box will look hot on a saliency map. Real visitors may skip it entirely.
What it means for a business website
A saliency map tends to over-predict faces, bright promo boxes, and images that real visitors pass over. If you use one to decide that your colorful offer panel is working, you may be wrong. If you use it to check whether your main button stands out more than the decoration around it, it can help.
What to do
- Use saliency maps to check first-glance visual weight, such as whether the primary CTA is the most prominent item in your hero.
- Do not use them to predict reading, scrolling, or conversions.
- When a saliency map and real visitor data disagree, trust the visitor data.
Sources
- Leiva et al.. Understanding Visual Saliency in Mobile User Interfaces. ACM MobileHCI, 2020.
- Bylinskii, Judd, Oliva, Torralba, Durand. What do different evaluation metrics tell us about saliency models?. IEEE TPAMI, 2019.
- Pernice. Banner Blindness Revisited: Users Dodge Ads on Mobile and Desktop. Nielsen Norman Group, 2018.